Orchestrating MCP

MCP Orchestration

This skill governs the strategic selection and execution of MCP tools across the entire factory ecosystem. It ensures that agents use the most efficient communication and retrieval patterns for the task at hand.

Process

The orchestration of MCP tools follows a strict sequence to maximize efficiency and minimize latency:

1. Tool Call Identification: Analyze the user request to identify which MCP tools are required (e.g., tavily_search, mcp_rag_search_library).
2. Skill-Level Memoization: Before execution, check if the deterministic call (e.g., a specific search query) is already cached. Apply the memoize_tool decorator from scripts/ai/memoization.py to handle this automatically.
3. Parallel Execution Strategy: Group independent tool calls. If searching for facts while reading local docs, combine them using the call_parallel function to execute concurrently.
4. Sequential Reasoning: For complex tasks, use mcp_sequential-thinking to plan the next steps based on the outputs of the parallel tool calls.
5. Failure Analysis & Retry: If a tool fails, invoke the Retry-with-Analysis pattern. Perform a diagnostic grounding step (e.g., list_dir) before adjusting parameters and retrying.

When to Use

This skill should be used when completing tasks related to orchestrating mcp.

Master Tool Reference (Fully Grounded)

1. Web Intelligence (Tavily)

• Search: mcp_tavily_tavily_search(query, search_depth, max_results) - Semantic web search.
• Research: mcp_tavily_tavily_research(input) - Multi-step research.
• Extract: mcp_tavily_tavily_extract(urls, query) - Clean markdown extraction.
• Map: mcp_tavily_tavily_map(url) - Domain sub-page discovery.
• Crawl: mcp_tavily_tavily_crawl(url) - Mass data harvesting.

2. Browser Automation (Playwright)

• Navigation: mcp_playwright_browser_navigate(url), mcp_playwright_browser_wait_for(text, time)
• Interaction: mcp_playwright_browser_click(ref), mcp_playwright_browser_type(ref, text), mcp_playwright_browser_drag(startRef, endRef), mcp_playwright_browser_hover(ref)
• Forms: mcp_playwright_browser_fill_form(fields), mcp_playwright_browser_select_option(ref, values), mcp_playwright_browser_file_upload(paths)
• State: mcp_playwright_browser_snapshot() (Accessibility tree), mcp_playwright_browser_take_screenshot(type)
• Analysis: mcp_playwright_browser_evaluate(function), mcp_playwright_browser_network_requests(includeStatic), mcp_playwright_browser_console_messages(level)
• Tabs: mcp_playwright_browser_tabs(action, index)

3. State & Knowledge (Memory & RAG)

• Memory Graph: mcp_memory_read_graph(), mcp_memory_create_entities(entities), mcp_memory_create_relations(relations), mcp_memory_add_observations(observations), mcp_memory_search_nodes(query), mcp_memory_open_nodes(names)
• Ebook Library (RAG): mcp_rag_search_library(query), mcp_rag_get_ebook_toc(document_name), mcp_rag_ingest_document(file_path), mcp_rag_list_library_sources()
• Repo Intelligence: mcp_deepwiki_ask_question(repoName, question), mcp_deepwiki_read_wiki_structure(repoName), mcp_deepwiki_read_wiki_contents(repoName)
• Framework Docs: mcp_docs-langchain_SearchDocsByLangChain(query)

4. Utilities (Fetch, Sequential Thinking, Parallelism)

• Fetch: mcp_fetch_fetch(url, raw, max_length) - Stateless HTTP retrieval.
• Reasoning: mcp_sequential-thinking_sequentialthinking(thought, thoughtNumber, totalThoughts, nextThoughtNeeded) - Structured problem solving.
• Parallelism: call_parallel(tool_calls: list) - Execute multiple MCP tools concurrently.

Workspace & Productivity Tools (Platform Specific)

5. Google Workspace (Gmail, Calendar, Contacts)

• Gmail: mcp_gmail_list_messages(query, max_results), mcp_gmail_get_message(message_id), mcp_gmail_create_draft(to, subject, body), mcp_gmail_send_message(to, subject, body)
• Calendar: mcp_google-calendar_calendar_list(), mcp_google-calendar_calendar_get_events(calendar_id, time_min, time_max), mcp_google-calendar_calendar_create_event(calendar_id, summary, start, end)
• Contacts: mcp_google-contacts_list_contacts(name_filter), mcp_google-contacts_search_contacts(query), mcp_google-contacts_get_contact(identifier), mcp_google-contacts_create_contact(given_name, family_name, email, phone)

6. Document & Professional Tools

• Doc Tools (Word): mcp_doc-tools_create_document(filename, content), mcp_doc-tools_edit_document(filename, actions), mcp_doc-tools_read_document(filename)
• GDrive: mcp_gdrive_list_files(query), mcp_gdrive_get_file_content(file_id)
• Excel: mcp_excel_write_data_to_excel(filepath, sheet_name, data), mcp_excel_read_data_from_excel(filepath, sheet_name)
• LangSmith: mcp_langsmith_list_projects(), mcp_langsmith_fetch_runs(project_name, limit, page_number)

Best Practices

• Grounding Priority: Memory > RAG > Workspace (if applicable) > Tavily/Fetch.
• Reasoning Integrity: Always use mcp_sequential-thinking_sequentialthinking before complex multi-agent handoffs.
• Cleanup: Delete temporary files and drafts after use.

Prerequisites

• Access to relevant project documentation
• Environmental awareness of the target stack